issue on testing lgcn

[eezxuan]

Hello,
We try to reproduce the result for your code of lgcn, but meet some issue. The detail is as follows:
1.We use the pretrained_lgcn_ms1m.pth from your model zoo, and your part1_test feature bin and meta, and the newest code in your github.
2. We first generate faiss_gpu_k_80.npz by scripts/tools/test_knn.sh, with --knn_method faiss_gpu
3. Then we run scripts/test_lgcn_ms1m.sh, with knn =80 and knn_method = 'faiss_gpu' in cfg_test_lgcn_ms1m.py
4. The resulting F-pairwise = 0.6958, F-bcubic = 0.7120. Which
5. If we change knn to 200, the result change to: f-pairwise = 0.18,F-bcubic =0.18

Thanks very much!

[yl-1993]

Hi @eezxuan , sorry for the late reply and thanks for reporting the issue.

Would you mind trying the provided faiss knns first? One potential problem is the inconsistent knn construction method between training and test. We have observed the performance drop when we use hnsw for training and use faiss for testing.

But we have not do not try your setting yet, we may try it first and report here later.

[ghost]

@eezxuan I have one small question. Which version of Pytorch and Cuda are you using to run the particular pretrained model. I am constantly having issues in Google Colab. Thanks in advance !

[yl-1993]

@redionxhepa Hi, for you knowledge, I used torch==0.5.0a0+e31ab99 and cuda 9.0 for experiments. Feel free to report issues encountered for different package versions.

[ghost]

I tried a part of the code and it works even in the new Colab Environments which I think have 1.x Pytorch versions and way more newer cuda versions. Thank you and nice job !